# Evolution 80 — Baseline Forensic Audit

FIRME AI CORPORATION · CAIN-42 · measured 2026-11-15 (UTC) · nothing here was deployed.

## 1. Repository state at the start

| Item | Observed |
|---|---|
| Repository root | `/root/cain-studio-platform` |
| Branch at start | `evolution-79` |
| HEAD at start | `f4449aa7913623bc034fa4434ca96c2a2919a889` |
| Reported baseline `f4449aa79` | exists; full hash `f4449aa7913623bc034fa4434ca96c2a2919a889` (committed 2026-11-15 15:53:43 UTC) |
| E78 final commit | `e2851428d2cd7ed9ae46f165e5d9b64c72632cf4`, an ancestor of the E79 commit (`git merge-base --is-ancestor`: true) |
| Commits between E78 final and E79 | 2: `67a0c74fc` (capability audit), `f4449aa79` (E79) |
| Tags | 253 (mostly `deploy/*`; `v42.84.0-aplusplus`, `cain-13/14-evolution-complete`). There is **no E79 or E78 tag**. |
| Remotes | three off-site bare repos (`149.28.193.50`, `207.246.66.130`, `45.76.169.191`); same administrative domain |
| Evolution branches | `evolution-47-48-49`, `evolution-69` … `evolution-79`; all preserved |
| Working tree | **2,166 uncommitted paths** (1,467 modified, 699 untracked) belonging to other sessions |
| Other worktrees | 7 (other sessions); untouched |

## 2. How Evolution 80 is isolated

The main working tree is shared by several concurrent sessions and is dirty. Evolution 80 therefore
runs in its own git worktree:

```
git worktree add -b evolution-80 /root/cain-e80-worktree f4449aa7913623bc034fa4434ca96c2a2919a889
```

The worktree started with **0 modified files**. No file in the shared tree was edited, staged or
committed. No other session's change can enter an Evolution 80 commit. Ownership claimed by this
evolution: `cain/cain2027/e80/**`, `scripts/evolution80/**`, `tests/test_cain2027_e80.py`,
`docs/evolution-80/**`, `artifacts/e80/**`, plus the single documented homepage correction listed in
`PUBLIC_EVIDENCE_RECONCILIATION.md`.

## 3. Production configuration (inspected, not changed)

- All three public names (`cainstudio.online`, `mcpgate.online`, `clawx.click`) resolve to this
  build host (`45.76.60.231`). The build host **is** the production host.
- Gateway `cain-platform-gateway` is active on port 8088. The second instance (port 8098) was
  **not listening** when probed.
- `audit.cainstudio.online` does not resolve.

### Production observation reported to the owner (not acted on)

At 16:13–16:17 UTC, `GET /` for `cainstudio.online` and `mcpgate.online` took **38 s or timed out**
(60 s), locally and publicly, while `/health` answered in 0.1 s and `clawx.click/` answered 200.
Host load average was 12–23 on 8 vCPU, driven by other workloads. Evolution 80 is prohibited from
modifying production, so nothing was restarted or changed. This needs the owner's attention.

## 4. Host and resources

| Resource | Build host | Second host `45.76.169.191` |
|---|---|---|
| vCPU | 8 (shared; load 12–23) | 2 |
| RAM | 30 GB (19 GB available) | 3.4 GB |
| GPU | none | none |
| Local model weights | Qwen2.5-0.5B-Instruct, Qwen2.5-1.5B-Instruct, SmolLM2-135M-Instruct | none; no torch |
| Sandboxes installed | bubblewrap only | — |
| llama.cpp / Ollama / vLLM / gVisor / Kata / Firecracker | not installed | not installed |

The second host cannot run model inference (2 vCPU, 3.4 GB, also hosts validator replicas). It is
used only for bundle re-verification.

## 5. Findings about the baseline

| ID | Finding | Evidence | Effect |
|---|---|---|---|
| F80-P0-1 | The three homepages carry **two** E79 blocks. The older v1 block is still labelled "EVOLUTION 79 LIVE" and shows v1 numbers under the v2 block. | `clawx-site/index.html`, `cainstudio-homepage/index.html`, `platform-gateway/frontend/index.html`; live `clawx.click` | Readers see 986 and 1,186 calls, 144/144 and 22/22, for "Evolution 79" on one page. See reconciliation. |
| F80-P0-2 | The E79 cybersecurity generator can produce only **18 distinct questions**. The E79 main set (seed 79) and replication set (seed 791) share **14 questions** (10 cybersecurity, 2 scientific, 1 software, 1 mathematics). | `tasks79.py`; recount in this audit | E79's replication was not fully independent of its main set. No model was trained on either, so no accuracy was inflated by training, but "independent replication" was overstated for that domain. |
| F80-P0-3 | The signed E79 v2 scorecard scores memory from "no-memory 0.0 → verified 0.0" (score 2.0) while the signed raw memory rows show **0.10 → 0.70**. | `e79_v2_master_summary.json` scorecard vs `memory_evaluation_v2.json` | The 5.78 composite (C-) was computed with a stale memory value. With the raw value the memory dimension would be 5.6 instead of 2.0. The grade is not rewritten; it is recorded as internally inconsistent. |
| F80-P0-4 | E79's "tool-augmented" arm lost label accuracy because the calculator protocol was applied to questions with nothing to compute. | E79 raw rows (`T_tool_single`: 14/104) | Confirmed; this is the design input for E80's applicability-aware routing. |
| F80-P0-5 | No E78/E79 git tag exists. | `git tag` | Baselines are identified by commit hash only. No tag was created by E80 (tagging a baseline is an owner decision). |

Historical evidence was not edited or deleted.
